International Journal of Science, Engineering and TechnologyAuthors: Research Scholar Viraj Raghunath Sonawane, Professor Balasaheb Agarkar, Associate Professor Sachin Chaudhari
Abstract: This review provides a comprehensive study of neural network-based perception pipelines for autonomous vehicles, concentrating on the accuracy-complexity trade-offs affecting Traffic Sign Recognition (TSR) and lane detecting systems. The survey outlines the progression from conventional vision algorithms to modern deep learning frameworks, emphasizing how Convolutional Neural Networks (CNNs) and analogous architectures have revolutionized robustness, feature discrimination, and real-time inference capabilities. In TSR, high-capacity architectures such as ResNet-101 and attention-augmented models demonstrate improved resilience to small-scale, occluded, and luminance-variable sign categories, while single-stage detection frameworks, exemplified by the YOLO family, offer latency-constrained inference suitable for embedded applications. Conversely, two-stage frameworks such as Faster R-CNN provide enhanced representational accuracy but entail much higher computational expenses, underscoring the intrinsic precision–efficiency trade-off characteristic of contemporary vision systems. This study broadens the inquiry into lane detection, where deep learning techniques encounter challenges related to environmental unpredictability, low-contrast features, occlusions, and sensor noise. Techniques employing structured ROI extraction significantly diminish unnecessary computation while preserving accuracy. The study combines algorithmic performance, computational constraints, and deployment viability to emphasize the importance of ongoing innovation in neural network topologies, as well as the provision of comprehensive and carefully annotated datasets for the unified approach of lane and traffic sign detection. Future research should focus on closing the gap between performance and the real time requirement of driver assistant systems, aiming for a fine mix of accuracy, efficiency, and reliability in dynamic real-world contexts.
